AE-UNet++: attention-enhanced UNet++ for building segmentation of remote sensing images

Abstract Semantic segmentation of high-resolution remote sensing images is crucial both theoretically and practically for building analysis. However, due to the structural complexity of buildings and the diversity of image backgrounds, traditional segmentation methods often struggle with issues such as boundary ambiguity, detail loss, and inadequate resolution. To address these issues, this paper proposes an enhanced AE-UNet++ network. We designed a Context Feature Fusion (CFF) module to address the problem of small or dispersed buildings in remote sensing images that are easily overlooked or blurred in standard convolution operations. This module effectively captures both global and local semantic information. Additionally, we introduced a Spatial Awareness Fusion (SAF) module at the network’s shortcut connections, enhancing the model’s ability to distinguish buildings from complex backgrounds. To further integrate and strengthen the representation of depth features across layers, we designed the Enhanced Convolution Block (ECB), which enables efficient fusion and enhancement of multi-level features. By combining advanced attention mechanisms with efficient feature fusion techniques, these innovative modules significantly improve the precision and robustness of building segmentation in high-resolution remote sensing images. Experimental results on the Massachusetts Building Dataset and WHU Building Dataset demonstrate that AE-UNet++ outperforms current mainstream segmentation methods.